# Geo-Redundancy for Azure AI Search

**Last updated:** 2026-02
**Status:** GA
**Category:** Business Continuity & Disaster Recovery

---

## Introduksjon

Azure AI Search er en regional tjeneste uten innebygd geo-replikering eller automatisk failover. Hvis regionen blir utilgjengelig, blir også search-tjenesten utilgjengelig. For AI-løsninger med RAG-arkitektur (Retrieval-Augmented Generation) er dette en kritisk svakhet fordi search-indeksen er hjørnesteinen i hele kunnskapsgjenfinningen.

For å oppnå geo-redundans for Azure AI Search må organisasjoner implementere egne løsninger: identiske search-tjenester i flere regioner, synkroniserte indekser, og load balancing med failover-logikk. Dette krever nøye planlegging av indekseringsstrategier, konsistensgarantier og trafikkstyring.

For norsk offentlig sektor med strenge tilgjengelighetskrav er multi-region AI Search en viktig komponent i BCDR-strategien. Typisk oppsett er primær i Norway East med sekundær i Sweden Central, noe som sikrer data residency innenfor EU/EØS samtidig som det gir regional redundans.

## Indeksreplikering på tvers av regioner

### Arkitekturoversikt

Azure AI Search har ingen innebygd mekanisme for indeksreplikering mellom regioner. Du må implementere en av følgende strategier:

```
Strategi 1: Dual Push Indexing
┌──────────────┐     ┌──────────────────┐     ┌──────────────────┐
│  Datakilde   │────▶│ Indexer Pipeline  │────▶│ Search Region A  │
│  (Blob/SQL)  │     │ (Azure Functions) │────▶│ Search Region B  │
└──────────────┘     └──────────────────┘     └──────────────────┘

Strategi 2: Pull from Replicated Source
┌──────────────┐     ┌──────────────────┐     ┌──────────────────┐
│ Datakilde A  │◀───▶│    GRS / GZRS    │◀───▶│ Datakilde B      │
│ (Region A)   │     │   Replikering    │     │ (Region B)       │
└──────┬───────┘     └──────────────────┘     └──────┬───────────┘
       │                                              │
       ▼                                              ▼
┌──────────────┐                              ┌──────────────────┐
│ AI Search A  │                              │ AI Search B      │
│ (Indexer)    │                              │ (Indexer)        │
└──────────────┘                              └──────────────────┘
```

### Dual Push Indexing med Azure Functions

```python
# Azure Function: Push-based dual-region indexing
import azure.functions as func
from azure.search.documents import SearchClient
from azure.core.credentials import AzureKeyCredential
import json

# Konfigurer klienter for begge regioner
primary_client = SearchClient(
    endpoint="https://search-primary-norwayeast.search.windows.net",
    index_name="knowledge-base",
    credential=AzureKeyCredential("<primary-key>")
)

secondary_client = SearchClient(
    endpoint="https://search-secondary-swedencentral.search.windows.net",
    index_name="knowledge-base",
    credential=AzureKeyCredential("<secondary-key>")
)

def main(msg: func.QueueMessage) -> None:
    """Process document and index to both regions."""
    document = json.loads(msg.get_body().decode('utf-8'))

    # Indekser til primær region
    try:
        primary_result = primary_client.upload_documents(documents=[document])
        logging.info(f"Primary indexed: {primary_result[0].key}")
    except Exception as e:
        logging.error(f"Primary indexing failed: {e}")
        # Send til dead-letter queue for retry
        raise

    # Indekser til sekundær region (asynkront er OK)
    try:
        secondary_result = secondary_client.upload_documents(documents=[document])
        logging.info(f"Secondary indexed: {secondary_result[0].key}")
    except Exception as e:
        logging.warning(f"Secondary indexing failed (will retry): {e}")
        # Legg i retry-kø — sekundær er ikke kritisk
        send_to_retry_queue(document)
```

### Pull-basert indeksering med Built-in Indexers

```bash
# Opprett identiske indexer i begge regioner
# Primær region — kobler til primær datakilde
az search indexer create \
  --service-name "search-primary-norwayeast" \
  --resource-group "rg-ai-prod" \
  --name "blob-indexer" \
  --data-source-name "blob-source-primary" \
  --target-index-name "knowledge-base" \
  --schedule '{"interval": "PT5M"}'

# Sekundær region — kobler til GRS-replikert datakilde
az search indexer create \
  --service-name "search-secondary-swedencentral" \
  --resource-group "rg-ai-dr" \
  --name "blob-indexer" \
  --data-source-name "blob-source-secondary" \
  --target-index-name "knowledge-base" \
  --schedule '{"interval": "PT5M"}'
```

## Replikatelling og dimensjonering for tilgjengelighet

### Intra-region tilgjengelighet med replikaer

Azure AI Search distribuerer automatisk replikaer på tvers av Availability Zones når du har 2+ replikaer i en region som støtter AZ.

| Replikaer | SLA | Lesbare spørringer | Skriveoperasjoner | Merknader |
|-----------|-----|-------------------|-------------------|-----------|
| 1 | 99.9% | Ja | Ja | Ingen AZ-redundans |
| 2 | 99.9% | Ja | Ja | AZ-distribuert automatisk |
| 3+ | 99.99% | Ja | Ja | Anbefalt for prod (read/write SLA) |

### Dimensjoneringsveiledning for multi-region

```
Per region (produksjon):
├── Replikaer: 3 (for 99.99% SLA og AZ-redundans)
├── Partisjoner: Basert på indeksstørrelse
│   ├── < 25 GB → 1 partisjon
│   ├── 25–50 GB → 2 partisjoner
│   ├── 50–150 GB → 3–6 partisjoner
│   └── > 150 GB → 6–12 partisjoner
└── SKU: Standard eller Standard S2/S3

Sekundær region (DR):
├── Replikaer: 2 (minimum for AZ, scale up ved failover)
├── Partisjoner: Identisk med primær
└── SKU: Identisk med primær
```

### Kostnadsoptimalisering for sekundær region

```bash
# Sekundær region starter med færre replikaer
# Scale up automatisk ved failover via Azure Automation

# Opprett Automation Runbook for scale-up
az automation runbook create \
  --automation-account-name "aa-ai-dr" \
  --resource-group "rg-ai-dr" \
  --name "scale-up-search-dr" \
  --type "PowerShell" \
  --content '
    # Scale sekundær AI Search fra 2 til 3 replikaer
    $searchService = Get-AzSearchService `
      -ResourceGroupName "rg-ai-dr" `
      -Name "search-secondary-swedencentral"
    Set-AzSearchService `
      -ResourceGroupName "rg-ai-dr" `
      -Name "search-secondary-swedencentral" `
      -ReplicaCount 3
    Write-Output "Scaled to 3 replicas for DR"
  '
```

## Failover- og routingstrategier

### Azure Front Door for AI Search failover

```bicep
// Bicep: Azure Front Door med failover for AI Search
resource frontDoor 'Microsoft.Cdn/profiles@2023-05-01' = {
  name: 'fd-ai-search'
  location: 'global'
  sku: {
    name: 'Premium_AzureFrontDoor'
  }
}

resource originGroup 'Microsoft.Cdn/profiles/originGroups@2023-05-01' = {
  parent: frontDoor
  name: 'search-origins'
  properties: {
    loadBalancingSettings: {
      sampleSize: 4
      successfulSamplesRequired: 3
    }
    healthProbeSettings: {
      probePath: '/indexes/knowledge-base/docs?api-version=2024-07-01&search=*&$top=1'
      probeRequestType: 'GET'
      probeProtocol: 'Https'
      probeIntervalInSeconds: 30
    }
  }
}

resource primaryOrigin 'Microsoft.Cdn/profiles/originGroups/origins@2023-05-01' = {
  parent: originGroup
  name: 'primary-norwayeast'
  properties: {
    hostName: 'search-primary-norwayeast.search.windows.net'
    priority: 1
    weight: 1000
  }
}

resource secondaryOrigin 'Microsoft.Cdn/profiles/originGroups/origins@2023-05-01' = {
  parent: originGroup
  name: 'secondary-swedencentral'
  properties: {
    hostName: 'search-secondary-swedencentral.search.windows.net'
    priority: 2
    weight: 1000
  }
}
```

### Application-level failover

```python
# Python: Application-level failover for Azure AI Search
from azure.search.documents import SearchClient
from azure.core.credentials import AzureKeyCredential
from azure.core.exceptions import ServiceResponseError, HttpResponseError
import time

class ResilientSearchClient:
    """AI Search client with automatic failover."""

    def __init__(self, primary_endpoint, secondary_endpoint, index_name, api_key):
        self.primary = SearchClient(
            endpoint=primary_endpoint,
            index_name=index_name,
            credential=AzureKeyCredential(api_key)
        )
        self.secondary = SearchClient(
            endpoint=secondary_endpoint,
            index_name=index_name,
            credential=AzureKeyCredential(api_key)
        )
        self.use_primary = True
        self.failover_time = None
        self.health_check_interval = 60  # sekunder

    def search(self, search_text, **kwargs):
        """Search with automatic failover."""
        client = self.primary if self.use_primary else self.secondary

        try:
            results = client.search(search_text=search_text, **kwargs)
            # Sjekk om vi kan falle tilbake til primær
            if not self.use_primary and self._should_check_primary():
                self._try_failback()
            return results

        except (ServiceResponseError, HttpResponseError) as e:
            if self.use_primary:
                print(f"Primary search failed, failing over: {e}")
                self.use_primary = False
                self.failover_time = time.time()
                return self.secondary.search(search_text=search_text, **kwargs)
            else:
                raise  # Begge regioner feiler

    def _should_check_primary(self):
        """Check if enough time has passed to try primary again."""
        if self.failover_time is None:
            return False
        return time.time() - self.failover_time > self.health_check_interval

    def _try_failback(self):
        """Attempt to fail back to primary region."""
        try:
            self.primary.search(search_text="*", top=1)
            self.use_primary = True
            self.failover_time = None
            print("Failback to primary successful")
        except Exception:
            pass  # Primær er fortsatt nede
```

## Holde indekser synkroniserte

### Synkroniseringstrategier

| Strategi | Forsinkelse | Kompleksitet | Anbefalt for |
|----------|-----------|-------------|-------------|
| Dual push (samtidige) | ~0 | Middels | Sanntidskritiske data |
| Event-driven sync | Sekunder | Middels | Generelt anbefalt |
| Scheduled indexer | 5–60 min | Lav | Batch-baserte oppdateringer |
| Full rebuild | Timer | Lav | Sjeldne endringer |

### Event-driven synkronisering med Event Grid

```bash
# Sett opp Event Grid for blob-endringer → trigger dual indexing
az eventgrid event-subscription create \
  --name "blob-change-to-search-sync" \
  --source-resource-id "/subscriptions/{sub}/resourceGroups/rg-ai-prod/providers/Microsoft.Storage/storageAccounts/staiprod" \
  --included-event-types "Microsoft.Storage.BlobCreated" "Microsoft.Storage.BlobDeleted" \
  --endpoint-type "azurefunction" \
  --endpoint "/subscriptions/{sub}/resourceGroups/rg-ai-prod/providers/Microsoft.Web/sites/func-search-sync/functions/SyncToSecondary"
```

### Indeks-konsistensvalidering

```python
# Periodisk validering av indekskonsistens mellom regioner
import requests

def validate_index_consistency(primary_endpoint, secondary_endpoint, index_name, api_key):
    """Compare document counts and sample documents between regions."""
    headers = {"api-key": api_key, "Content-Type": "application/json"}

    # Sammenlign dokumenttellinger
    primary_count = requests.get(
        f"{primary_endpoint}/indexes/{index_name}/docs/$count?api-version=2024-07-01",
        headers=headers
    ).json()

    secondary_count = requests.get(
        f"{secondary_endpoint}/indexes/{index_name}/docs/$count?api-version=2024-07-01",
        headers=headers
    ).json()

    drift = abs(primary_count - secondary_count)
    drift_pct = (drift / max(primary_count, 1)) * 100

    return {
        "primary_count": primary_count,
        "secondary_count": secondary_count,
        "drift": drift,
        "drift_percentage": round(drift_pct, 2),
        "in_sync": drift_pct < 1.0  # < 1% avvik er akseptabelt
    }
```

## Query-ytelse i multi-region oppsett

### Latensoptimalisering

| Strategi | Latensreduksjon | Merknad |
|----------|----------------|---------|
| Latency-based routing | 20–50 ms | Brukere sendes til nærmeste region |
| Semantic caching | 80–95% | Cache hyppige spørringer i APIM |
| Read replicas (intra-region) | 10–30 ms | Fordel lesninger over replikaer |
| Query optimalisering | Varierer | $select, $top for å redusere payload |

### Azure API Management for caching og routing

```xml
<!-- APIM Policy: Cache og failover for AI Search queries -->
<policies>
  <inbound>
    <base />
    <cache-lookup vary-by-query-parameter="search,filter,top,skip"
                  caching-type="internal" />
  </inbound>
  <backend>
    <retry condition="@(context.Response.StatusCode >= 500)"
           count="1" interval="0" first-fast-retry="true">
      <choose>
        <when condition="@(context.Variables.GetValueOrDefault<bool>("usePrimary", true))">
          <set-backend-service
            base-url="https://search-primary-norwayeast.search.windows.net" />
        </when>
        <otherwise>
          <set-backend-service
            base-url="https://search-secondary-swedencentral.search.windows.net" />
        </otherwise>
      </choose>
    </retry>
  </backend>
  <outbound>
    <cache-store duration="300" />
    <base />
  </outbound>
</policies>
```

## Referanser

- [Multi-region deployments in Azure AI Search](https://learn.microsoft.com/en-us/azure/search/search-multi-region) — Offisiell multi-region guide
- [Reliability in Azure AI Search](https://learn.microsoft.com/en-us/azure/reliability/reliability-ai-search) — Tilgjengelighet, AZ og DR
- [Multi-region solutions in nonpaired regions](https://learn.microsoft.com/en-us/azure/reliability/regions-multi-region-nonpaired) — Multi-region uten parede regioner
- [Azure AI Search multi-region Bicep sample](https://github.com/Azure-Samples/azure-search-multiple-regions) — Komplett Bicep-mal
- [Azure Front Door overview](https://learn.microsoft.com/en-us/azure/frontdoor/front-door-overview) — Global load balancing
- [Azure Traffic Manager overview](https://learn.microsoft.com/en-us/azure/traffic-manager/traffic-manager-overview) — DNS-basert trafikkstyring

## For Cosmo

- **Bruk denne referansen** når kunden bygger RAG-løsninger med Azure AI Search og trenger geo-redundans for search-indeksene.
- Azure AI Search har INGEN innebygd geo-replikering — dette er en vanlig misforståelse. Kunden må implementere dual-indexing selv.
- Anbefal minimum 3 replikaer per region for 99.99% SLA og AZ-redundans — 2 replikaer gir kun 99.9%.
- For kostnadsoptimalisering: Sekundær region kan kjøre med 2 replikaer og skalere opp til 3 ved failover via Azure Automation.
- Indekskonsistens bør valideres automatisk — sett opp periodisk sjekk av dokumenttelling og samplingsbasert innholdsvalidering.